Energy-Efficient CPU Core Allocation in DVFS and CPU Pinning Enabled Cloud

Improving the energy efficiency of servers in data centers is one of the critical issues in cloud computing. A promising technology to address this issue is dynamic voltage and frequency scaling (DVFS). Since CPU core frequency affects power consumption more than proportionally, multi-core CPU sin DVFS can minimize overall CPU power consumption by equalizing the CPU core load for distributing their frequencies to lower values. To equalize the CPU core load of servers in cloud data centers, where multiple services share CPU resources, an appropriate number of CPU cores must be allocated to services. This paper proposes a CPU core allocation method for minimizing CPU power consumption in a cloud that assumes DVFS and CPU pinning. This method dynamically allocates the appropriate number of CPU cores to services on the basis of the variance of CPU core frequencies. The major advantage of the method is its general applicability to a variety of services and server platforms. Also, it can be combined with various DVFS technologies since it only optimizes the allocation of the CPU cores and does not directly manipulate the CPU cores' frequencies. Experiments using a web server showed that the proposed method reduces CPU power consumption by up to 25% compared with when CPU cores required for performance are allocated, as is common in the conventional cloud. Experiments also showed that the proposed method is effective for various CPUfreq governors, which are mainstream implementations of DVFS.